Data Information Granular-based Multi-Scale Prediction And Scheduling For Industrial Energy System

With high energy consumption and serious pollution in industrial production,it has become a major national demand to improve the utilization efficiency of industrial energy and realize energy conservation and emission reduction.With a large amount of energy generation and consumption in the process of industrial production,the effective prediction and optimal scheduling of energy medium play a vital role to achieve energy saving and emission reduction as well as ensure a safe production status.Considering that the existing data point-based modeling methods cannot appropriately describe the production semantics or equipment operating characteristics,plus their analysis scale is limited,some data granular-based multiscale prediction and scheduling methods are proposed in this study as followsRegarding multi-time scale prediction of industrial energy media,an adaptive granulationbased prediction model and a featural granularity-based long-term prediction model are proposed respectively.The first one partitions and describes the information granules according to the semantic features of the industrial data,and uses collaborate-conditional clustering and fuzzy inference to achieve prediction.Aimming at the long-term prediction demand,the second one further feeds the granular features into a multivariate long short-term memory network so that the feature-temporal pattern can be utilized to make predictions.To enhance long-term prediction performance,a periodic correction method is also adopted to reduce the iteration error caused by multi-step prediction process.In view of the spatio-temporal series prediction of industrial energy media,a granular spatio-temporal topology model based on long short-term memory networks is proposed.The spatio-temporal data are first partitioned according to production process and represented into featural spatial granular series,and a multi-level long short-term memory network is established by considering the spatio-temporal topology order of the granular series.In order to reduce the model complexity,a gating mechanism is further put forward within the unit structure,with which the network can understand the time-order of spatial granules and adjust the updating and prediction procedures accordingly.Aiming at the scheduling problem of industrial energy system,a granular prediction and dynamic scheduling process based on adaptive dynamic programming is proposed.Considering the slow adjustment process and the event-driven characteristic,the accumulated reward of critic network is calculated and the scheduling action network is established in unit of information granules.In order to estimate the future trends of energy storage levels to real-time determine the scheduling moments,a reinforcement learning-based semi-supervised granulation method is proposed to perform factor prediction process.As for the long-term scheduling problem of industrial energy system,a granular modeling and reinforcement learning-based method is proposed.The human experience-based initial policies are calculated based on granular partition and fuzzy inference.To achieve multi-step dynamic scheduling,a two-stage value function estimation method is proposed in the critic part,where the state transition process of the system is depicted by a granular prediction model to simulate the dynamic changes of the production environment when taking scheduling actions.Finally,the dynamic compensation for the initial policies are performed by the actor part.Utilizing practical operation data from the energy center of a steel company in China,the experiments on the above prediction and scheduling methods are thoroughly conducted.The results show that the prediction accuracy of the proposed method can meet the on-site requirements for long-term prediction,and has advantages in the prediction of some key production processes within spatio-temporal relationship.By combining the granular prediction process with the scheduling model,the proposed method can provide safe and economical scheduling schemes for the applied system.Owing to the consideration of the dynamic characteristics of production environment and the influence of multi-step events,the proposed method can deliver effective solutions for long-term scheduling scenarios.